Support Hellinger-distance-based scoring both on SDAR 1D and 2D

[takuti]

What this PR includes

By keeping the last _mu and _sigma in the SDAR algorithm as _muOld and _sigmaOld, we are now able to compute the Hellinger distance between the models before/after update. Usage is demonstrated as comments in ChangeFinder1D and ChangeFinder2D.

Computation of the Hellinger distance between two (multivariate) normal distributions are implemented on MathUtils.

Are the Hellinger distances computed correctly?

Unit tests have not been implemented yet. However, I have confirmed whether it works as expected by comparing the results of Python implementation.

• mu1: 2.25635999045
• sigma1: 1.56789907203
• mu2: 2.37802392246
• sigma2: 4.5496514822

For the above parameters, the Python code returns a distance 0.0661275442145.

Here, our MathUtils.hellingerDistanes() method also works as:

MathUtils.hellingerDistance(2.25635999045, 1.56789907203, 2.37802392246, 4.5496514822);
// => almost equal to 0.0661275442145

RealVector mu1 = new ArrayRealVector(new double[] {2.25635999045});
RealMatrix sigma1 = new Array2DRowRealMatrix(new double[] {1.56789907203});
RealVector mu2 = new ArrayRealVector(new double[] {2.37802392246});
RealMatrix sigma2 = new Array2DRowRealMatrix(new double[] {4.5496514822});
MathUtils.hellingerDistance(mu1, sigma1, mu2, sigma2);
// => almost equal to 0.0661275442145

[myui]

if(denominator == 0.d) {
return 1.d;
}

[myui]

Merged. Thanks. I'll fix some.